Research map / Continual learning / Online and replay-based
Continual Learning for Motion Prediction Model via Meta-Representation Learning and Optimal Memory Buffer Retention Strategy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a continual motion prediction (CMP) framework that combines OML-style sparse meta-representation pre-training (with an L1 modulation network) and a distribution-shift-aware adaptive memory buffer, evaluated on nuScenes against joint training, fine-tuning, OML, MAML-Rep and EWC.
Problem
Motion prediction models for autonomous driving degrade in new driving environments, and continually updating them suffers from high retraining cost (joint training) and catastrophic forgetting (fine-tuning / transfer learning) when data arrives in small batches.
Approach
Pre-train a representation network with OML-style meta-learning plus an L1-regularized modulation network that induces sparse representations, and during adaptation keep a memory buffer updated only when the representation similarity between incoming data and the buffer falls below a threshold (distribution-shift detection) rather than by reservoir sampling.
Key results
- With memory buffer size 40, the proposed method reaches RA 2.06±0.11 and BTI 0.13±0.09 (ADE), lower than standard supervised (2.42 / 0.16), OML, EWC and MAML-Rep.
- Proposed buffer sampling yields 11% dead neurons versus 24% for reservoir sampling and 31% for random sampling.
- Pre-trained representation without update achieves RA 2.57 vs 4.96 for standard supervised pre-training, and its training time per adaptation stage grows more slowly than joint training (Fig. 8).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed pre-training and adaptation framework instead of standard supervised pre-training | Motion prediction model · nuScenes continual update, M=20 | RA (ADE): 2.87±0.11 -> 2.24±0.23 | Helped |
| Use the proposed pre-training and adaptation framework instead of standard supervised pre-training | Motion prediction model · nuScenes continual update, M=30 | RA (ADE): 2.89±0.52 -> 2.16±0.31 | Helped |
| Use the proposed pre-training and adaptation framework instead of standard supervised pre-training | Motion prediction model · nuScenes continual update, M=40 | RA (ADE): 2.42±0.08 -> 2.06±0.11 | Helped |
| Add adaptive sampling to reservoir sampling for memory-buffer retention (ablation) | Continual motion prediction adaptation · nuScenes adaptation, M=20 | RA (ADE): 2.54±0.25 -> 2.45±0.12 | Helped |
| Add adaptive sampling to reservoir sampling for memory-buffer retention (ablation) | Continual motion prediction adaptation · nuScenes adaptation, M=30 | RA (ADE): 2.54±0.08 -> 2.48±0.49 | Helped |
| Add adaptive sampling to reservoir sampling for memory-buffer retention (ablation) | Continual motion prediction adaptation · nuScenes adaptation, M=40 | RA (ADE): 2.48±0.14 -> 2.25±0.13 | Helped |
| Add memory replay using random sampling instead of no memory replay (ablation) | Scratch continual adaptation · nuScenes adaptation, M=20 | RA (ADE): 3.94±1.09 -> 3.42±0.51 | Helped |
| Replace random sampling with reservoir sampling for memory replay (ablation) | Scratch continual adaptation · nuScenes adaptation, M=20 | RA (ADE): 3.42±0.51 -> 2.54±0.25 | Helped |
| Add sparse representation learning with a modulation network (ablation) | Pre-training for motion prediction · nuScenes, without model update | RA (ADE): 3.24±0.19 -> 2.57±0.09 | Helped |
| Use the proposed method instead of OML pre-training (ablation) | Motion prediction model · nuScenes, without model update | RA (ADE): 3.24±0.19 -> 2.57±0.09 | Helped |
| Use the proposed method instead of standard supervised pre-training (ablation) | Motion prediction model · nuScenes, without model update | RA (ADE): 4.96±0.13 -> 2.57±0.09 | Helped |
| Increase the memory buffer size | Proposed continual motion prediction method · nuScenes continual update | RA (ADE) | Helped |
Surprising findings
- Standard supervised pre-training shows higher RA than the proposed method at buffer sizes 20 and 30 in Table 1 region, so gains are buffer-size dependent.
- Joint training yields a single-number ADE of 1.92 and is reported without a BTI, indicating the comparison is not on identical metrics.
Limitations
- Evaluated only on nuScenes with a virtual continual scenario built from scene splits, not on real streaming deployment.
- Reported gains are small relative to variance in several settings and the paper provides no code or released artifacts.
- Pre-training and adaptation hyperparameters and sparsity regularization weight details are sparse.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: ADE retained accuracy (RA) backward-transfer and interference (BTI) training time dead neuron percentage instance sparsity
Baselines: Joint training Standard supervised pre-training OML MAML-Rep EWC Scratch Random sampling Reservoir sampling
Builds on: Online-aware Meta-Learning (OML) reservoir sampling Elastic Weight Consolidation MAML
Closest papers by meaning
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.85)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.84)
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.83)
- Semi-supervised Class-Agnostic Motion Prediction with Pseudo Label Regeneration and BEVMix (AAAI 2024 · similarity 0.83)
- Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation (AAAI 2024 · similarity 0.82)
- C-NAV: Towards Self-Evolving Continual Object Navigation in Open World (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.